Presentation is loading. Please wait.

Presentation is loading. Please wait.

Deep Learning Packages

Similar presentations


Presentation on theme: "Deep Learning Packages"— Presentation transcript:

1 Deep Learning Packages
with Daniel L. Silver, Ph.D. Christian Frey, BBA April 11-12, 2017

2 Benefits of a Deep Learning Framework
Handles the background tasks of weight update, gradient descent, loss calculation Can generate weight shapes automatically based on input/output/layer sizes. Allows easy use of GPU acceleration or lower level language acceleration (ex. Use C for matrix multiplication when in Python)

3 Keras Framework on top of TensorFlow or Theano
Follows the principle of layers – can stack, split or merge for unique network architectures. Calculates the connection size between hidden layers based on each layers size. Allows GPU acceleration with minimal configuration.

4 Theano Came out of the University of Montreal and Yoshua Bengio’s lab
Function based, many calls go through theano.function([variables], function), function can be any user defined function (c = a + b) Easy use of Numpy and GPU acceleration for integration with your current programs Python based Learn more at

5 Torch A core neural network library with many add-on packages for different tasks, such as image processing and visualization Allows you to build arbitrary graph networks (like TensorFlow), and run them on either CPU or GPU Originally created by developers from Facebook, Google Deepmind, and Twitter Uses LuaJIT, a just in time version of Lua that is several times faster than standard Lua

6 DeepLearning4J As the name suggests, a deep learning library for Java, Scala, and Clojure. Development led by Skymind.ai Allows for integration with Keras to import models built in Python. The Arbiter subproject works to find ideal metaparameters for you. Supports GPUs and CPU’s with optimized matrix processing in C++ and BLAS support.

7 Caffe Came out of the Berkley Vision and Learning Center (BVLC)
Designed for speed and modularity, and can process over 60M images per day on 1 GPU. Focuses on Computer Vision and convolution networks Has binding for Python and MATLAB Supports GPUs for faster training, as well as OpenBLAS for faster matrix operations. Found at


Download ppt "Deep Learning Packages"

Similar presentations


Ads by Google